$ cat briefs/daily/2026-11-08.md
2026-11-08
November 8, 2026 (Sun)
1 science evaluation · 1 hosted-agent release · 2 paper picks
Top Stories
1. BioBasicsBench tests whether agents respect experimental evidence
Across 100 routine wet-lab tasks, OpenAI reports 66.7% mean reward for GPT-6 Astra at xhigh, versus 43.8% for GPT-5.6 Sol. These are shares of available rubric credit, not complete-task success rates. The report still finds assumptions overriding evidence.
Why it matters: better average scores do not establish reliable lab judgments; an evaluation must also test when evidence is insufficient. → OpenAI (source)
2. Kimi Hosted Agents makes the runtime a managed service
Moonshot's 2026-11-05 announcement describes isolated sandboxes, network controls, interruptible sessions and audit exports. Its separate pre-inference security hook is currently offered selectively to enterprise-verified users.
Why it matters: runtime controls become part of the product, but the announcement does not independently establish isolation or productivity gains. → Moonshot AI · Claude Managed Agents (source)
Paper Picks
Harness-S: rescue failed training attempts with sibling harnesses, then return the trained policy to its original harness. Read it for separating training assistance from deployment dependencies; the abstract reports improvements without numerical effect sizes. → Harness-S (source)
Hybrid Computer Use: combine code, accessibility and graphical interaction, with protection against stale-state actions during channel switching. Read it for interface-aware evaluation; reported OSWorld 2.0 gains have no numerical scores in the abstract. First submitted 2026-10-30. → Hybrid Computer Use (source)
Watch
- Harness-S's inference router adds gains at equal executor rollout count. Total routing and training cost remains an open question. → Agentic Reinforcement Learning (source)
New in Wiki
- Both paper pages above add authors, methods and evidence limits; no new entity, concept or person page needs review. → Harness-S · Hybrid Computer Use (training) (interface)
Updates
- Eval Harness Configuration records Arena's 2026-11-04 release: Claude Opus 5.5 (Max), 20.14% Net Improvement, 95% CI 18.06% to 22.23%. This is a signed contrast against a randomized baseline, not Elo or success probability. (source)